You open an API task, and the skill you left installed has already taken a slice of your context. A token is the unit your AI plan bills by, about three quarters of a word. The context is everything the model can see while it answers you. Among these skills, 11,602 load their names and descriptions every session.
The load decision behind backend skills
A skill is a folder of instructions the model opens when it decides it needs them. A glob is a filename pattern, like *.tsx. The pattern can appear in instructions, but it does not decide when the skill loads.
Keep general guidance small. Call deeper help for a specific API or database task. The other 423 wait until you invoke them.
That split matters when you install several skills for one project. A focused guide such as api-and-interface-design can shape API boundaries without making every database question carry a large body of instructions.
What backend skills cost before you need them
The median preload for every session is 58 tokens. At 20 sessions a day, that becomes 1,160 tokens across a day. The 90th percentile is 143 tokens every session, so a few large descriptions can dominate a bundle.
The biggest listed preload is atproto-lexicon, at 720 tokens every session. That is closer to a page of instructions than a quick hint. On-demand skills have a median body of 1,022 tokens, but that body waits until the model loads it.
What the backend scan flags
The scan graded 11,530 skills A, 296 B, 198 C or D, and 1 E or F. Those are counts of skills, not token costs. It recorded 1,716 network-call findings, 190 findings that send data to an external URL, and 84 that run shell commands.
The A-grade x402 still has a low-severity network-call finding. The D-grade rest-graphql-debug is listed with external data sending, remote-code execution, and network calls. add-whatsapp is listed at grade C with recursive force deletion.
Every finding is printed with the line that produced it, giving you something concrete to inspect. The scan reports what it saw in the file; it does not decide whether the skill suits your project.
Backend skills worth starting with
The data tail says how much arrives before work, how long the file is, and how many copies the catalogue knows. The original is the first copy of a file we saw; everything matching it afterwards is a copy.
- studio-mock-api-tests: A testing guide for Supabase Studio components that intercepts network requests with MSW. 79 tokens every session, 298 lines, original, 0 copies.
- azure-openai-to-responses: A guide for moving Python Azure OpenAI apps from Chat Completions to the Responses API. 185 tokens every session, 162 lines, original, 0 copies.
- dce-edge: It keeps Node-only code out of edge builds through safer conditional imports. 84 tokens every session, 69 lines, original, 0 copies.
- blocking-io-guard: It finds blocking input and output in asynchronous Python backend code. 111 tokens every session, 142 lines, original, 0 copies.
- build-mcp-server: plans and builds an MCP server, the program that exposes your tools and data to an agent. 111 tokens every session, 222 lines, original, 11 copies.
- api-and-interface-design: It helps design stable REST or GraphQL interfaces and boundaries between software parts. 49 tokens every session, 368 lines, original, 30 copies.
When a hook beats a backend skill
Choose a hook when you need an action tied to an event, such as a check after a change. A hook is a shell command your agent runs for you when something happens.
A skill supplies guidance after it is selected, so it cannot replace event scheduling. The scan recorded 84 “Runs shell commands” findings, but that does not turn those skills into hooks.
Questions people ask
Are these skills for Claude Code or Cursor? Of these backend skills, 11,477 are labelled for Claude Code, 6,137 for Codex and 146 for Cursor, and one skill can carry more than one label. Check the target agent before installing.
Should I install a whole plugin for one API task? Only if you need most of what it ships. Installing a plugin installs every skill it contains, so unused preload still counts.
How large can a full API skill become when loaded? The Azure OpenAI migration guide opens a body of 2,035 tokens. Its 185-token preload is separate, so leaving it available costs less than opening all its instructions.
Do it now
Run npx agentmods add <slug> for the skill you choose, then compare the backend catalogue at Skills.